Data Augmentation for Private Posterior Estimation

A data augmentation based sampler for conducting privacy-aware Bayesian inference. The dapper_sample() function takes an existing sampler as input and automatically constructs a privacy-aware sampler. The process of constructing a sampler is simplified through the specification of four independent modules, allowing for easy comparison between different privacy mechanisms by only swapping out the relevant modules. Probability mass functions for the discrete Gaussian and discrete Laplacian are provided to facilitate analyses dealing with privatized count data. The output of dapper_sample() can be analyzed using many of the same tools from the 'rstan' ecosystem. For methodological details on the sampler see Ju et al. (2022) , and for details on the discrete Gaussian and discrete Laplacian distributions see Canonne et al. (2020) .

dapper

A data augmentation framework for privacy aware Bayesian inference.

Installation

Install the development version, including its vignette, with:

devtools::install_github("mango-empire/dapper", build_vignettes = TRUE)

Example workflow

The randomized-response vignette walks through model construction, posterior sampling, diagnostics, and a comparison with an analysis that ignores privacy noise. After installation, open it with:

vignette("randomized-response", package = "dapper")